When I do multiclass segmentation and use DiceMulti/ JaccardCoeffMulti, they always return an average value, while I want to view certain value for certain organ. How can I modify it? So much obliged for anyone answering it.
I’m not very familiar with segmentation, so someone else with more experience may know of a built-in way to get what you need, but here’s my hacky attempt which I think gets you the right output. You can see the full code in this Colab notebook.
In it, I redefine DiceMulti (all I do is change the last line from return np.nanmean(binary_dice_scores) to return binary_dice_scores so it returns all of the scores and not just the mean)—not sure if that’s what you are looking for.
1 Like
Thank you so much for your help. Though that link is not accessible, you description inspired me=).
I write a CustomDiceMulti with class_index as parameter.
class CustomDiceMulti(Metric):
"Averaged Dice metric (Macro F1) for multiclass target in segmentation"
def __init__(self, axis=1, class_index=None):
self.axis = axis
self.class_index = class_index
def reset(self): self.inter,self.union = {},{}
def accumulate(self, learn):
pred,targ = flatten_check(learn.pred.argmax(dim=self.axis), learn.y)
for c in range(learn.pred.shape[self.axis]):
p = torch.where(pred == c, 1, 0)
t = torch.where(targ == c, 1, 0)
c_inter = (p*t).float().sum().item()
c_union = (p+t).float().sum().item()
if c in self.inter:
self.inter[c] += c_inter
self.union[c] += c_union
else:
self.inter[c] = c_inter
self.union[c] = c_union
@property
def value(self):
# Get intermediate calculations
inter, union = self.get_inter_union()
if self.class_index is not None:
# Calculate and return for the specified class
return 2 * inter[self.class_index] / union[self.class_index] if union[self.class_index] > 0 else np.nan
else:
# Calculate mean across classes
binary_dice_scores = np.array([])
for c in inter:
binary_dice_scores = np.append(binary_dice_scores, 2 * inter[c] / union[c] if union[c] > 0 else np.nan)
return np.nanmean(binary_dice_scores)
def get_inter_union(self):
"""Returns the intersection and union counts for each class."""
return self.inter, self.union
It may help someone but I suggest second check on the math thing. And thank you again for it really helping me.
1 Like
Oh interesting! I wasn’t sure what inter was but now that makes sense.
Here’s the full code of DiceMulti that I edited—your approach is better at getting the values directly though:
class DiceMulti(Metric):
"Averaged Dice metric (Macro F1) for multiclass target in segmentation"
def __init__(self, axis=1): self.axis = axis
def reset(self): self.inter,self.union = {},{}
def accumulate(self, learn):
pred,targ = flatten_check(learn.pred.argmax(dim=self.axis), learn.y)
for c in range(learn.pred.shape[self.axis]):
p = torch.where(pred == c, 1, 0)
t = torch.where(targ == c, 1, 0)
c_inter = (p*t).float().sum().item()
c_union = (p+t).float().sum().item()
if c in self.inter:
self.inter[c] += c_inter
self.union[c] += c_union
else:
self.inter[c] = c_inter
self.union[c] = c_union
@property
def value(self):
binary_dice_scores = np.array([])
for c in self.inter:
binary_dice_scores = np.append(binary_dice_scores, 2.*self.inter[c]/self.union[c] if self.union[c] > 0 else np.nan)
# return np.nanmean(binary_dice_scores)
return binary_dice_scores
this is the output during training (messy but at least shows the values):
And then .recorder.log gets the array binary_dice_scores for the last epoch:
And here’s my Colab link trying again: Google Colab
I see and thank you again for your help. 
1 Like